Magnetic resonance imaging (MRI) simulations necessitate a substantial number of multi-contrast MR images, which can be time-consuming and costly to obtain. To overcome this challenge, synthetic data has emerged as a viable alternative. However, existing methods are limited in their ability to generate multiple modalities within a single dataset and struggle to produce modalities that are not present in the original dataset, resulting in poor domain transfer capability. To tackle this issue, we propose MCGAN for cross-domain multi-contrast MR image synthesis. The MCGAN framework employs a two-stage learning strategy. In the first stage, a domain adaptation module is utilized to align the source domain distribution with the target domain using unsupervised learning, effectively bridging domain gaps. Subsequently, the second stage involves an image-to-image module that empowers the model to generate additional modalities. By combining these two stages, MCGAN framework overcomes the limitations of single-stage generation methods, resulting in a model that synthesizes a comprehensive array of modalities within each dataset. Experimental results show that MCGAN method outperforms other transfer learning-based image-to-image methods and cross-dataset image synthesis methods in terms of both data distribution realism and texture details. Code is available at https://github.com/DropInOcean/MCGAN .